[X] I am a LangChain maintainer, or was asked directly by a LangChain maintainer to create an issue here.
Issue Content
Context: Open (meaning weights-available here) models have come a long way! When it comes to reliable tool use over multiple turns, they still require more programming (via better instructions, few-shots, etc.) than most of the leading-edge models (gpt-4*, Opus, Gemini even). Because of this, the maintainers of this repo haven't yet taken the time to develop an excellent, reliable reference implementation using Llama 3, Mixtral, or another "open" model that anyone can use and it "just works".
People are especially interested in the Multi-agent supervisor pattern and in customer support-type use-cases where some agentic loop is incorporated.
We'd love to make better examples of common design patterns and add evals to demonstrate where they work more broadly to make it easier for folks to build more advanced applications without having to use closed APIs.
These examples must be of high quality. If no one gets to it, I'll try to implement myself, but I haven't currently allocated cycles for this.
Privileged issue
Issue Content
Context: Open (meaning weights-available here) models have come a long way! When it comes to reliable tool use over multiple turns, they still require more programming (via better instructions, few-shots, etc.) than most of the leading-edge models (gpt-4*, Opus, Gemini even). Because of this, the maintainers of this repo haven't yet taken the time to develop an excellent, reliable reference implementation using Llama 3, Mixtral, or another "open" model that anyone can use and it "just works".
People are especially interested in the Multi-agent supervisor pattern and in customer support-type use-cases where some agentic loop is incorporated.
We'd love to make better examples of common design patterns and add evals to demonstrate where they work more broadly to make it easier for folks to build more advanced applications without having to use closed APIs.
These examples must be of high quality. If no one gets to it, I'll try to implement myself, but I haven't currently allocated cycles for this.